Text Mining the EMR for Modeling and Predicting Suicidal Behavior among US Veterans of the 1991 Persian Gulf War

Abstract

Suicide is an important public health problem and prominent among U.S. Veterans and active duty troops. Prediction of suicide and suicide attempts is problematic because these are low-frequency events and traditional clinical screening approaches have a high false positive rate. Large clinical databases extracted from electronic health records permit study of suicidal behavior in larger populations than previously possible using sampling techniques. In addition to offering structured data, text search and classification methods can identify additional risk variables. Data extracted from clinical records of 250,000 veterans were modeled using machine learning methodology. To predict suicide attempts in this population over a 10 year period. In contrast to previously reported models, our results showed high specificity and a false positive rate of 0.5%, contrasting with other studies showing false positive rates exceeding 20%. The model showed lower specificity with a true positive rate of 27% and a false negative rate of 73%. These results suggest that a machine learning approach developed with large data sets can usefully supplement current approaches to prediction of suicidal behavior.